Professional Documents
Culture Documents
Xiaogang Wang
Outline
Introduction
Shape quantification
Compactness
Spatial moments
Radial distance measures
Fourier descriptors
Thinning
Texture quantification
Statistical moments
Co-occurrence matrix measures
Spectral measures
2
Tissue shape,
Ti
h
size,
i texture
t t
andd density
d it
Musculoskeletal angle, kinematics, and stress
Ventricular motion
Myocardial strain
Blood flow
Tumor growth
Cytometric features
It is
i often
ft based
b d image
i
segmentation
t ti results
lt
3
Shape Quantification
FF
benign
g
Malignant
Translation invariant
Scale invariant
Rotation invariant
Compactness
P is the perimeter
A is the area
Normalize compactness C ' 1 4 / C
It increases with increasing shape complexity
Visual shape roughness perception often may have a good correlation
with C,
C but not always the case
Spatial Moments
Moments pq x0 y 0 x y f ( x, y ) x, y 0,1,2,3
They provide an effective set of shape descriptors
p+q is the moment order
pq x 0
where
m10
m00
N 1
(x
y0
x ) p ( y y ) q f ( x, y ) x, y 0,1,2,3
and
m01
m00
10
Spatial Moments
11
Spatial Moments
02
i 20
2
411
( 20 02 ) 2
Spatial
p
moments have been used to qquantify
y breast tumors,,
mammographic calcifications and blood cells.
12
(xN-1, yN-1)
(x1, y1)
(x2, y2)
(x , y )
c
(xn, yn)
13
Statistical moments
Discrete Fourier transform
14
15
16
17
18
19
20
21
s(p
( 1) is
i the
th sum off 0-to-1
0 t 1 transitions
t
iti
in
i the
th sequence off
neighbor pixels, computed by considering all nine transitions
around the pixel p1, starting with p2 to p3 and end with p9 to p2.
22
Condition 1: 2 n(p1) 6. If p1 has only one neighbor, it is the end of a pixel strand
and should not be deleted
deleted. If p1 has 7 neighbors
neighbors, deleting it may deplete the region
and lead to splitting.
Condition 2: s(p1) =1. If the neighborhood has more than one 0-to-1 transition,
delete p1 mayy lead to splitting
p
g the region.
g
Condition 3: p2 * p4 * p6 = 0
Condition 4: p4 * p6 * p8 = 0.
These two conditions are satisfied simultaneously if p4 = 0, or p6 = 0, or p2 = p8 = 0.
When conditions 1 and 2 are met, these three possibilities refer to the three cases
illustrated in the figure below
If all the four conditions are met, p1 is marked for deletion but not deleted until all
the pixels are evaluated. After the complete evaluation, all marked pixels are
deleted.
East border
23
Condition 3: p2 * p4 * p8 = 0
Condition 4: p2 * p6 * p8 = 0
These two conditions address the cases of north border, west border, and southeast
corner in a manner similar to step 1.
If all the four conditions are met, p1 is marked for deletion but not deleted until all
the pixels are evaluated. After the complete evaluation, all marked pixels are
deleted.
When step
p 2 is completed,
p
the algorithm
g
starts another iteration by
y
applying step 1 to the outcome of step 2.
(a) Small section of an image. (b) Outcome of thinning with the Zhang
and Suen algorithm, black pixels remain and gray pixels are removed.
25
(a)
(b)
(a) Thinning with the MAT. (b) Thinning with the Zhang and Suen algorithm
26
(a) An atomic force microscope image of DNA fragments. (b) Image after thresholding. (c)
Image after thinning with Zhang and Suen algorithm and removing objects that are too short
or touching the edge of the image.
27
28
A normal vein
A vein exhibiting
g beading
g
29
Texture Quantification
30
Ultrasound image sections of normal liver (a), fatty liver (b) and
liver with cirrhosis (c)
31
Visualized using
the Lambertian
reflection model
Nevus
melanoma
32
Statistical Moments
Th dont
They
d encode
d spatial
i l information
i f
i
33
Co-occurrence Matrix
34
Co-occurrence Matrix
Example
Image region
Normalized co-occurrence
co occurrence matrix C(d,
C(d ),
) cij hij / ij hij
If K is too large, need to remap the gray values using contrast
enhancement techniques and re-quantize gray values to a
smaller number of gray levels.
35
36
37
38
Applications
39
Spectral Measures
40
Spectral Measures
F*(u,
( , v)) is the complex
p conjugate
j g of F(u,
( , v))
41
Spectral Measures
Spectral Measures
43
Reading
44